Long-Term Memory

chimera.sessions.long_term_memory provides persistent memory that survives across agent sessions. Store facts, learnings, preferences, and project context that the agent can recall later. Backed by a JSON file for simplicity.

Inspired by Codex’s memory consolidation system.

Key Classes

Class	Description
LongTermMemory	Key-value memory store with categories, search, and prompt rendering
MemoryEntry	A single entry with key, content, category, timestamps, and metadata

Quick Start

from chimera.sessions.long_term_memory import LongTermMemory

memory = LongTermMemory("~/.chimera/memory.json")

memory.store("user_name", "Alice", category="preference")
memory.store("project_lang", "Python 3.12", category="project")
memory.store("test_framework", "pytest", category="project")

# In a later session:
memory = LongTermMemory("~/.chimera/memory.json")
name = memory.recall("user_name")  # "Alice"

Categories and Search

# Recall all memories in a category
project = memory.recall_category("project")

# Search by substring (case-insensitive, matches key and content)
results = memory.search("python")
for entry in results:
    print(f"{entry.key}: {entry.content}")

Prompt Injection

to_prompt_section() renders memories as Markdown for injection into system prompts:

section = memory.to_prompt_section(categories=["preference", "project"])
# Returns:
# ## Agent Memory
# ### Preference
# - **user_name**: Alice
# ### Project
# - **project_lang**: Python 3.12

Managing Memories

memory.forget("old_key")   # Remove a single entry (returns True/False)
memory.clear()              # Remove all entries
print(memory.count)         # Number of stored memories

Import Reference

from chimera.sessions.long_term_memory import LongTermMemory, MemoryEntry

Related

• Sessions — multi-turn conversation persistence
• Checkpoints — named context snapshots